Good find!

I don't think we're getting the constraints quite right though - and unfortunately I'm worried that we may need to bake the label adjustment in at a deeper level, as constrained autorange now happens inside concatExtremes:

That's probably only part of the solution though, as that's just for matches constraints. scaleanchor constraints are handled in constraints.enforce:

axes_chain_scaleanchor_matches2_inside-ticklabels.png perhaps shows the issue most clearly: one of the features of axes_chain_scaleanchor_matches2.png that's really useful is comparing successive rows, green vs red, and black vs blue. The left-most red & green subplots should be identical, as should the right-most blue and black, and in both cases the domain-constrained subplots should still fill the available space (dotted rectangle), because within the subplot there's only a matches constraint, which cannot reduce a domain. Yet both the left-most green and right-most black in the inside version have constrained domains, and in the green one this even clips one of the data points.

From one subplot to the next there's a scaleanchor constraint, so subsequent green subplots (and earlier black subplots) should end up with smaller domains, which we do see in the new one. But in axes_chain_scaleanchor_matches2.png the data points in each green subplot are positioned exactly where the red data points are, and same blue vs black, which I believe should still be the case for the inside case, which mostly looks borne out in the inside case, but not for the first green subplot.

ticklabelposition3.png and axes_chain_scaleanchor_matches_inside-ticklabels.png have their x labels overlapping the image / heatmaps a bit, which I suspect is a manifestation of the same issue.